The Signal
Photonic Quantum Model Reproduces Multivariate Correlations
A preprint describes a continuous-variable photonic circuit trained to match correlations in financial data, with affine processing used to adjust individual marginals. The authors also report results on 20 variables and a separate method for simulating weak Kerr nonlinearities.

Dat Chi Le and Oleksandr Kyriienko propose a photonic continuous-variable model for learning correlated multivariate data. Their preprint, submitted to arXiv on 3 November 2026, uses homodyne detection on a photonic circuit and trains the circuit to match correlations measured in data. The paper frames the approach as a way to generate samples from distributions whose variables have structured relationships.
How the circuit shapes correlations
The model uses an N-mode photonic circuit, with each mode contributing a variable to the output. Two-mode squeezing creates correlations between modes. Local phase shifts tune the strength and sign of those correlations, and a programmable interferometer redistributes them across mode pairs. The authors describe the roles of these circuit components in terms of their effects on the output correlation matrix.
Training uses the Pearson correlation matrix from the data as its target. Rather than training by matching samples, the circuit is trained classically by matching the covariance of the simulated state to the empirical correlation structure. The paper says this makes it possible to train using the state’s covariance alone. Homodyne detection of the trained circuit then supplies continuous-valued samples.
After fitting the correlations, the authors apply affine post-processing to adjust the marginal distributions independently while keeping the relationships intact. This separates the adjustment of individual variables’ marginals from the circuit’s task of representing correlations. The model is therefore assessed both on the structure linking variables and on its ability to accommodate different marginal distributions.
Test on financial returns
Le and Kyriienko test the architecture on the multivariate distribution of standardized log returns for financial assets. They examine circuits with up to twenty modes. The paper reports that, for five variables, the model reproduces empirical correlations with residuals below 10−4.
For twenty variables, the authors report that the correct signs are recovered for 94% ± 1% of 190 correlation pairs, with 86% of the residuals below 0.1. The sign result and residual figures measure different aspects of the fit: whether the model gets the direction of each pairwise correlation right, and how closely its correlations match the empirical values. The abstract presents these as results on the financial-return distribution, not as a guarantee for other datasets.
Separate work on Kerr gates
The preprint also develops an approach for simulating Kerr gates with small nonlinearity. It represents the weak-strength action through a Hubbard-Stratonovich expansion, which the authors say makes the gates suitable for continuous-variable-based simulators. This is a technical contribution alongside the correlation-learning model; it is distinct from the financial-return evaluation.
The paper places the model among generative approaches that use physical systems to produce samples. Its proposed circuit uses Gaussian operations to form and distribute correlations, and homodyne measurements to obtain continuous outputs. The authors also include a Kerr nonlinearity to move beyond the Gaussian family, while representing its weak-strength action in a way that can be handled with Gaussian phase-space methods. They present the resulting circuits as a platform for interpretable generative learning and for representing and fast sampling of distributions across application domains.
Source
Related stories
- Photonic Processor Classifies Quantum States by Stellar Rank
- Hybrid Quantum Model Tracks Fermion Scattering Near Resonance
- Qluster raises €4.7 million for photonic quantum computing
- ORCA PT-2 data yields larger malware-model refinement gains in study
- Quandela Names Quang Trung University for Vietnam’s First Quantum Computer